Error: Could not load the stable-diffusion model! Reason: Weights only load failed.

[Stugist]

Bug occurs on image generation.
Fresh linux install, no warnings in shell other than one saying that local:9000 is already populated.
With no changes made, default "Astronaut on a Horse" I get this error:

Error: Could not load the stable-diffusion model! Reason: Weights only load failed. This file can still be loaded, to do so you have two options, �[1mdo those steps only if you trust the source of the checkpoint�[0m. (1) In PyTorch 2.6, we changed the default value of the weights_only argument in torch.load from False to True. Re-running torch.load with weights_only set to False will likely succeed, but it can result in arbitrary code execution. Do it only if you got the file from a trusted source. (2) Alternatively, to load with weights_only=True please check the recommended steps in the following error message. WeightsUnpickler error: Unsupported global: GLOBAL pytorch_lightning.callbacks.model_checkpoint.ModelCheckpoint was not an allowed global by default. Please use torch.serialization.add_safe_globals([ModelCheckpoint]) or the torch.serialization.safe_globals([ModelCheckpoint]) context manager to allowlist this global if you trust this class/function. Check the documentation of torch.load to learn more about types accepted by default with weights_only https://pytorch.org/docs/stable/generated/torch.load.html.

Now, I read through as much of the documentation at the end as I could understand. Which frankly is next to nothing.
I have no idea what's happening, or even where to find an error log.

[Stugist]

I restarted my laptop and ran start.sh again, the following is the output from that.
On the first few attempts running ED there were no warnings in the terminal, as stated above.
the 497 lines preceding the below consists of many many many variations of:

cp: cannot remove .......file path......: permission denied

/run/media/geist/Skinny/easy-diffusion/installer_files/env/lib/python3.9/site-packages/conda/base/context.py:202: FutureWarning: Adding 'defaults' to channel list implicitly is deprecated and will be removed in 25.3.

To remove this warning, please choose a default channel explicitly with conda's regular configuration system, e.g. by adding 'defaults' to the list of channels:

conda config --add channels defaults

For more information see https://docs.conda.io/projects/conda/en/stable/user-guide/configuration/use-condarc.html

deprecated.topic(
torchruntime: 1.15.0
torch: 2.6.0+cu124
torchvision: 0.21.0+cu124
sdkit: 2.0.22.8
stable-diffusion-sdkit: 2.1.5
diffusers: 0.28.2
{'model': {'stable-diffusion': 'sd-v1-4'},
'models_dir': '/run/media/geist/Skinny/easy-diffusion/models',
'net': {'listen_port': 9000, 'listen_to_network': False},
'render_devices': ['cuda'],
'ui': {'open_browser_on_start': True},
'update_branch': 'main',
'use_v3_engine': False,
'vram_usage_level': 'balanced'}

Easy Diffusion installation complete, starting the server!

--- GPUs ---

0. GPU(vendor_id='8086', vendor_name='Intel Corporation', device_id='9bc4', device_name='CometLake-H GT2 [UHD Graphics]', is_discrete=False)
1. GPU(vendor_id='10de', vendor_name='NVIDIA Corporation', device_id='1f99', device_name='TU117M [GeForce GTX 1650 Mobile / Max-Q]', is_discrete=True)

--- RECOMMENDED TORCH PLATFORM ---
cu124

--- CONFIGURATION ---
PYTHONPATH=/run/media/geist/Skinny/easy-diffusion/installer_files/env/lib/python3.8/site-packages
Python: /run/media/geist/Skinny/easy-diffusion/installer_files/env/bin/python
Version: 3.9.21
Checking network settings
Set listen port to 9000

Launching uvicorn

20:10:06.223 INFO MainThread started in server.py:31
/run/media/geist/Skinny/easy-diffusion/stable-diffusion
20:10:06.227 INFO MainThread started at 03/08/25 20:10:06 server.py:32
/run/media/geist/Skinny/easy-diffusion/installer_files/env/lib/python3.9/site-packages/pydantic/_internal/_config.py:345: UserWarning: Valid config keys have changed in V2:

• 'orm_mode' has been renamed to 'from_attributes'
  warnings.warn(message, UserWarning)
  stable-diffusion model(s) found.
  gfpgan model(s) found.
  realesrgan model(s) found.
  vae model(s) found.
  20:10:06.583 INFO MainThread Start new Rendering Thread on device: cuda task_manager.py:410
  20:10:06.589 INFO cuda Device usage during initialization: runtime.py:35
  20:10:06.591 INFO cuda CPU utilization: 0.0%, System RAM used: 3.7 of 15.5 GiB, GPU memory_utils.py:58
  RAM used (cuda:0): 0.1 of 3.6 GiB (peak: 0.0 GiB)
  20:10:06.592 WARNING cuda forcing full precision on this GPU, to avoid corrupted device_manager.py:120
  images. GPU: NVIDIA GeForce GTX 1650
  20:10:06.593 INFO cuda Setting cuda as active, with precision: full device_manager.py:123
  20:10:06.600 INFO cuda loading stable-diffusion model from init.py:52
  /run/media/geist/Skinny/easy-diffusion/models/stable-diffusion/sd-v1-4.ckpt to device:
  cuda
  No module 'xformers'. Proceeding without it.
  20:10:07.622 INFO MainThread Opening browser.. app.py:317
  ╭────────────────────────────────────── Easy Diffusion is ready ──────────────────────────────────────╮
  │ │
  │ Easy Diffusion is ready to serve requests. │
  │ │
  │ A new browser tab should have been opened by now. │
  │ If not, please open your web browser and navigate to http://localhost:9000/ │
  │ │
  ╰─────────────────────────────────────────────────────────────────────────────────────────────────────╯
  20:10:08.795 INFO AnyIO worker thread Scanning all model folders for models... model_manager.py:418
  20:10:08.882 ERROR cuda Error while loading stable-diffusion model: model_manager.py:82
  /run/media/geist/Skinny/easy-diffusion/models/stable-diffusion/sd-v1-4.ckpt
  20:10:08.885 ERROR cuda Weights only load failed. This file can still be loaded, to model_manager.py:88
  do so you have two options, [1mdo those steps only if you trust the source of the
  checkpoint[0m.
  (1) In PyTorch 2.6, we changed the default value of the weights_only
  argument in torch.load from False to True. Re-running torch.load with
  weights_only set to False will likely succeed, but it can result in arbitrary
  code execution. Do it only if you got the file from a trusted source.
  (2) Alternatively, to load with weights_only=True please check the
  recommended steps in the following error message.
  WeightsUnpickler error: Unsupported global: GLOBAL
  pytorch_lightning.callbacks.model_checkpoint.ModelCheckpoint was not an allowed
  global by default. Please use
  torch.serialization.add_safe_globals([ModelCheckpoint]) or the
  torch.serialization.safe_globals([ModelCheckpoint]) context manager to allowlist
  this global if you trust this class/function.

Check the documentation of torch.load to learn more about types accepted by default
with weights_only https://pytorch.org/docs/stable/generated/torch.load.html.
Traceback (most recent call last):
File "/run/media/geist/Skinny/easy-diffusion/ui/easydiffusion/model_manager.py",
line 73, in load_default_models
load_model(
File
"/run/media/geist/Skinny/easy-diffusion/installer_files/env/lib/python3.9/site-pack
ages/sdkit/models/model_loader/init.py", line 54, in load_model
context.models = get_loader_module(model_type).load_model(context, **kwargs)
File
"/run/media/geist/Skinny/easy-diffusion/installer_files/env/lib/python3.9/site-pack
ages/sdkit/models/model_loader/stable_diffusion/init.py", line 85, in
load_model
sd = load_tensor_file(model_path)
File
"/run/media/geist/Skinny/easy-diffusion/installer_files/env/lib/python3.9/site-pack
ages/sdkit/utils/file_utils.py", line 15, in load_tensor_file
return torch.load(path, map_location="cpu")
File
"/run/media/geist/Skinny/easy-diffusion/installer_files/env/lib/python3.9/site-pack
ages/torch/serialization.py", line 1470, in load
raise pickle.UnpicklingError(_get_wo_message(str(e))) from None
_pickle.UnpicklingError: Weights only load failed. This file can still be loaded,
to do so you have two options, [1mdo those steps only if you trust the source of
the checkpoint[0m.
(1) In PyTorch 2.6, we changed the default value of the weights_only
argument in torch.load from False to True. Re-running torch.load with
weights_only set to False will likely succeed, but it can result in arbitrary
code execution. Do it only if you got the file from a trusted source.
(2) Alternatively, to load with weights_only=True please check the
recommended steps in the following error message.
WeightsUnpickler error: Unsupported global: GLOBAL
pytorch_lightning.callbacks.model_checkpoint.ModelCheckpoint was not an allowed
global by default. Please use
torch.serialization.add_safe_globals([ModelCheckpoint]) or the
torch.serialization.safe_globals([ModelCheckpoint]) context manager to allowlist
this global if you trust this class/function.

Check the documentation of torch.load to learn more about types accepted by default
with weights_only https://pytorch.org/docs/stable/generated/torch.load.html.
20:10:09.297 INFO AnyIO worker thread Scanned 3 models. Nothing infected model_manager.py:429
20:13:16.639 INFO cuda Session 1741464608395 starting task 126436485205008 on task_manager.py:283
NVIDIA GeForce GTX 1650
20:13:16.680 INFO cuda loading stable-diffusion model from init.py:52
/run/media/geist/Skinny/easy-diffusion/models/stable-diffusion/sd-v1-4.ckpt to device:
cuda
20:13:18.843 ERROR cuda Weights only load failed. This file can still be loaded, model_manager.py:187
to do so you have two options, [1mdo those steps only if you trust the source of
the checkpoint[0m.
(1) In PyTorch 2.6, we changed the default value of the weights_only
argument in torch.load from False to True. Re-running torch.load with
weights_only set to False will likely succeed, but it can result in arbitrary
code execution. Do it only if you got the file from a trusted source.
(2) Alternatively, to load with weights_only=True please check the
recommended steps in the following error message.
WeightsUnpickler error: Unsupported global: GLOBAL
pytorch_lightning.callbacks.model_checkpoint.ModelCheckpoint was not an allowed
global by default. Please use
torch.serialization.add_safe_globals([ModelCheckpoint]) or the
torch.serialization.safe_globals([ModelCheckpoint]) context manager to allowlist
this global if you trust this class/function.

Check the documentation of torch.load to learn more about types accepted by
default with weights_only
https://pytorch.org/docs/stable/generated/torch.load.html.
Traceback (most recent call last):
File "/run/media/geist/Skinny/easy-diffusion/ui/easydiffusion/model_manager.py",
line 183, in reload_models_if_necessary
action_fn(context, model_type, scan_model=False, **extra_params) # we've
scanned them already
File
"/run/media/geist/Skinny/easy-diffusion/installer_files/env/lib/python3.9/site-pac
kages/sdkit/models/model_loader/init.py", line 54, in load_model
context.models = get_loader_module(model_type).load_model(context, **kwargs)
File
"/run/media/geist/Skinny/easy-diffusion/installer_files/env/lib/python3.9/site-pac
kages/sdkit/models/model_loader/stable_diffusion/init.py", line 85, in
load_model
sd = load_tensor_file(model_path)
File
"/run/media/geist/Skinny/easy-diffusion/installer_files/env/lib/python3.9/site-pac
kages/sdkit/utils/file_utils.py", line 15, in load_tensor_file
return torch.load(path, map_location="cpu")
File
"/run/media/geist/Skinny/easy-diffusion/installer_files/env/lib/python3.9/site-pac
kages/torch/serialization.py", line 1470, in load
raise pickle.UnpicklingError(_get_wo_message(str(e))) from None
_pickle.UnpicklingError: Weights only load failed. This file can still be loaded,
to do so you have two options, [1mdo those steps only if you trust the source of
the checkpoint[0m.
(1) In PyTorch 2.6, we changed the default value of the weights_only
argument in torch.load from False to True. Re-running torch.load with
weights_only set to False will likely succeed, but it can result in arbitrary
code execution. Do it only if you got the file from a trusted source.
(2) Alternatively, to load with weights_only=True please check the
recommended steps in the following error message.
WeightsUnpickler error: Unsupported global: GLOBAL
pytorch_lightning.callbacks.model_checkpoint.ModelCheckpoint was not an allowed
global by default. Please use
torch.serialization.add_safe_globals([ModelCheckpoint]) or the
torch.serialization.safe_globals([ModelCheckpoint]) context manager to allowlist
this global if you trust this class/function.

Check the documentation of torch.load to learn more about types accepted by
default with weights_only
https://pytorch.org/docs/stable/generated/torch.load.html.
20:13:18.924 ERROR cuda Traceback (most recent call last): task_manager.py:295
File "/run/media/geist/Skinny/easy-diffusion/ui/easydiffusion/task_manager.py",
line 287, in thread_render
task.run()
File
"/run/media/geist/Skinny/easy-diffusion/ui/easydiffusion/tasks/render_images.py",
line 89, in run
model_manager.fail_if_models_did_not_load(context)
File "/run/media/geist/Skinny/easy-diffusion/ui/easydiffusion/model_manager.py",
line 216, in fail_if_models_did_not_load
raise Exception(f"Could not load the {model_type} model! Reason: " + e)
Exception: Could not load the stable-diffusion model! Reason: Weights only load
failed. This file can still be loaded, to do so you have two options, [1mdo those
steps only if you trust the source of the checkpoint[0m.
(1) In PyTorch 2.6, we changed the default value of the weights_only
argument in torch.load from False to True. Re-running torch.load with
weights_only set to False will likely succeed, but it can result in arbitrary
code execution. Do it only if you got the file from a trusted source.
(2) Alternatively, to load with weights_only=True please check the
recommended steps in the following error message.
WeightsUnpickler error: Unsupported global: GLOBAL
pytorch_lightning.callbacks.model_checkpoint.ModelCheckpoint was not an allowed
global by default. Please use
torch.serialization.add_safe_globals([ModelCheckpoint]) or the
torch.serialization.safe_globals([ModelCheckpoint]) context manager to allowlist
this global if you trust this class/function.

Check the documentation of torch.load to learn more about types accepted by default
with weights_only https://pytorch.org/docs/stable/generated/torch.load.html.

[jpa-54]

Today : one install on Linux laptop without GPU and another install on Linux Desktop with Nvidia GPU. In this two cases, sames issues with same error when I try to make image: "Error: Could not load the stable-diffusion model! Reason: Weights only load failed.........."

Best regards

[basdp]

Same here, Linux OpenSUSE Tumbleweed x86_64 with NVIDIA RTX 3090 Ti and latest NVIDIA proprietary drivers.

[kellyj4]

Not sure this is a bug but more of a change in behaviour in PyTorch.

If you are happy with a workaround you can set the following ENV var in Mac or Linux before running start.sh

export TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1

This worked for me today on a Mac but as the warning message states there maybe security concerns in doing this.

[jpa-54]

Not sure this is a bug but more of a change in behaviour in PyTorch.

If you are happy with a workaround you can set the following ENV var in Mac or Linux before running start.sh

export TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1

This worked for me today on a Mac but as the warning message states there maybe security concerns in doing this.

OK fine ! its works for me. Thanks !

[cmdr2]

Hi @basdp @kellyj4 @jpa-54 @Stugist - Thanks! I'd suggest downloading and using the safetensors format model instead of this one: https://github.com/easydiffusion/sdkit-test-data/releases/download/assets/sd-v1-4.safetensors

I've also fixed the installer. On new installations, it will now download a safetensors model instead of ckpt.

Please let me know if this continues to be an issue. Thanks!

[ugintl]

Not sure this is a bug but more of a change in behaviour in PyTorch.

If you are happy with a workaround you can set the following ENV var in Mac or Linux before running start.sh

export TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1

This worked for me today on a Mac but as the warning message states there maybe security concerns in doing this.

I am on windows. Exactly where do I have to put it?